基于分类激活图增强的立体视觉图像分类方法

文章摘要

基于分类激活图增强的立体视觉图像分类方法

Stereo vision image classification method based on classification activation map enhancement

投稿时间：2022-11-16 修订日期：2022-11-16

DOI：

中文关键词: 分类激活图像增强立体视觉图像双线性特征卷积神经网络

英文关键词: Classification activated image enhancement Stereo vision image Bilinear feature Convolution neural network

基金项目:2021年福建省中青年教师教育科研项目基于暗通道的运动图像复原算法研究（JAT210595）

作者	单位	邮编
吴清平^*	闽南理工学院信息管理学院	362700

摘要点击次数: 123

全文下载次数: 0

中文摘要:

在立体视觉图像分类过程中，因无法准确计算出图像双线性特征的通道维数，导致图像分类准确率较低。为此，提出基于分类激活图增强的立体视觉图像分类方法。利用卷积神经网络对提取图像双线性特征，通过外积运算将双线性汇合特征展开成一维特征，获取特征通道维数，通过特征间元素的累加和交互达到特征融合的目的。以融合结果为基础，对图像特征进行细化，利用分类激活图增强方法对细化特征的待分类图像进行池化处理，实现立体视觉图像的分类。测试结果表明，所提方法立体视觉图像分类损失较小，分类准确率较高。

英文摘要:

In the process of stereo vision image classification, the accuracy of image classification is low because the channel dimension of image bilinear features cannot be calculated accurately. Therefore, a stereo vision image classification method based on classification activation map enhancement is proposed. The convolution neural network pair is used to extract the bilinear features of the image, and the bilinear confluence features are expanded into one-dimensional features through the outer product operation to obtain the dimension of feature channels. The feature fusion is achieved through the accumulation and interaction of elements between features. Based on the fusion results, the image features are refined, and the classification activation map enhancement method is used to pool the images to be classified with the refined features to achieve the classification of stereo vision images. The test results show that the proposed method has less classification loss and higher classification accuracy.

View Fulltext 查看/发表评论下载PDF阅读器

关闭